Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Device operations for the pnfs nfs4 file layout driver.
4 *
5 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
6 *
7 * Tao Peng <bergwolf@primarydata.com>
8 */
9
10#include <linux/nfs_fs.h>
11#include <linux/vmalloc.h>
12#include <linux/module.h>
13#include <linux/sunrpc/addr.h>
14
15#include "../internal.h"
16#include "../nfs4session.h"
17#include "flexfilelayout.h"
18
19#define NFSDBG_FACILITY NFSDBG_PNFS_LD
20
21static unsigned int dataserver_timeo = NFS_DEF_TCP_RETRANS;
22static unsigned int dataserver_retrans;
23
24static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg);
25
26void nfs4_ff_layout_put_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
27{
28 if (!IS_ERR_OR_NULL(mirror_ds))
29 nfs4_put_deviceid_node(&mirror_ds->id_node);
30}
31
32void nfs4_ff_layout_free_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
33{
34 nfs4_print_deviceid(&mirror_ds->id_node.deviceid);
35 nfs4_pnfs_ds_put(mirror_ds->ds);
36 kfree(mirror_ds->ds_versions);
37 kfree_rcu(mirror_ds, id_node.rcu);
38}
39
40/* Decode opaque device data and construct new_ds using it */
41struct nfs4_ff_layout_ds *
42nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
43 gfp_t gfp_flags)
44{
45 struct xdr_stream stream;
46 struct xdr_buf buf;
47 struct page *scratch;
48 struct list_head dsaddrs;
49 struct nfs4_pnfs_ds_addr *da;
50 struct nfs4_ff_layout_ds *new_ds = NULL;
51 struct nfs4_ff_ds_version *ds_versions = NULL;
52 u32 mp_count;
53 u32 version_count;
54 __be32 *p;
55 int i, ret = -ENOMEM;
56
57 /* set up xdr stream */
58 scratch = alloc_page(gfp_flags);
59 if (!scratch)
60 goto out_err;
61
62 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
63 if (!new_ds)
64 goto out_scratch;
65
66 nfs4_init_deviceid_node(&new_ds->id_node,
67 server,
68 &pdev->dev_id);
69 INIT_LIST_HEAD(&dsaddrs);
70
71 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
72 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
73
74 /* multipath count */
75 p = xdr_inline_decode(&stream, 4);
76 if (unlikely(!p))
77 goto out_err_drain_dsaddrs;
78 mp_count = be32_to_cpup(p);
79 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
80
81 for (i = 0; i < mp_count; i++) {
82 /* multipath ds */
83 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
84 &stream, gfp_flags);
85 if (da)
86 list_add_tail(&da->da_node, &dsaddrs);
87 }
88 if (list_empty(&dsaddrs)) {
89 dprintk("%s: no suitable DS addresses found\n",
90 __func__);
91 ret = -ENOMEDIUM;
92 goto out_err_drain_dsaddrs;
93 }
94
95 /* version count */
96 p = xdr_inline_decode(&stream, 4);
97 if (unlikely(!p))
98 goto out_err_drain_dsaddrs;
99 version_count = be32_to_cpup(p);
100 dprintk("%s: version count %d\n", __func__, version_count);
101
102 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
103 gfp_flags);
104 if (!ds_versions)
105 goto out_scratch;
106
107 for (i = 0; i < version_count; i++) {
108 /* 20 = version(4) + minor_version(4) + rsize(4) + wsize(4) +
109 * tightly_coupled(4) */
110 p = xdr_inline_decode(&stream, 20);
111 if (unlikely(!p))
112 goto out_err_drain_dsaddrs;
113 ds_versions[i].version = be32_to_cpup(p++);
114 ds_versions[i].minor_version = be32_to_cpup(p++);
115 ds_versions[i].rsize = nfs_block_size(be32_to_cpup(p++), NULL);
116 ds_versions[i].wsize = nfs_block_size(be32_to_cpup(p++), NULL);
117 ds_versions[i].tightly_coupled = be32_to_cpup(p);
118
119 if (ds_versions[i].rsize > NFS_MAX_FILE_IO_SIZE)
120 ds_versions[i].rsize = NFS_MAX_FILE_IO_SIZE;
121 if (ds_versions[i].wsize > NFS_MAX_FILE_IO_SIZE)
122 ds_versions[i].wsize = NFS_MAX_FILE_IO_SIZE;
123
124 /*
125 * check for valid major/minor combination.
126 * currently we support dataserver which talk:
127 * v3, v4.0, v4.1, v4.2
128 */
129 if (!((ds_versions[i].version == 3 && ds_versions[i].minor_version == 0) ||
130 (ds_versions[i].version == 4 && ds_versions[i].minor_version < 3))) {
131 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
132 i, ds_versions[i].version,
133 ds_versions[i].minor_version);
134 ret = -EPROTONOSUPPORT;
135 goto out_err_drain_dsaddrs;
136 }
137
138 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
139 __func__, i, ds_versions[i].version,
140 ds_versions[i].minor_version,
141 ds_versions[i].rsize,
142 ds_versions[i].wsize,
143 ds_versions[i].tightly_coupled);
144 }
145
146 new_ds->ds_versions = ds_versions;
147 new_ds->ds_versions_cnt = version_count;
148
149 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
150 if (!new_ds->ds)
151 goto out_err_drain_dsaddrs;
152
153 /* If DS was already in cache, free ds addrs */
154 while (!list_empty(&dsaddrs)) {
155 da = list_first_entry(&dsaddrs,
156 struct nfs4_pnfs_ds_addr,
157 da_node);
158 list_del_init(&da->da_node);
159 kfree(da->da_remotestr);
160 kfree(da);
161 }
162
163 __free_page(scratch);
164 return new_ds;
165
166out_err_drain_dsaddrs:
167 while (!list_empty(&dsaddrs)) {
168 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
169 da_node);
170 list_del_init(&da->da_node);
171 kfree(da->da_remotestr);
172 kfree(da);
173 }
174
175 kfree(ds_versions);
176out_scratch:
177 __free_page(scratch);
178out_err:
179 kfree(new_ds);
180
181 dprintk("%s ERROR: returning %d\n", __func__, ret);
182 return NULL;
183}
184
185static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
186 struct nfs4_deviceid_node *devid)
187{
188 nfs4_delete_deviceid(devid->ld, devid->nfs_client, &devid->deviceid);
189 if (!ff_layout_has_available_ds(lseg))
190 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
191 lseg);
192}
193
194static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
195 struct nfs4_ff_layout_mirror *mirror,
196 bool create)
197{
198 if (mirror == NULL || IS_ERR(mirror->mirror_ds))
199 goto outerr;
200 if (mirror->mirror_ds == NULL) {
201 if (create) {
202 struct nfs4_deviceid_node *node;
203 struct pnfs_layout_hdr *lh = lseg->pls_layout;
204 struct nfs4_ff_layout_ds *mirror_ds = ERR_PTR(-ENODEV);
205
206 node = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
207 &mirror->devid, lh->plh_lc_cred,
208 GFP_KERNEL);
209 if (node)
210 mirror_ds = FF_LAYOUT_MIRROR_DS(node);
211
212 /* check for race with another call to this function */
213 if (cmpxchg(&mirror->mirror_ds, NULL, mirror_ds) &&
214 mirror_ds != ERR_PTR(-ENODEV))
215 nfs4_put_deviceid_node(node);
216 } else
217 goto outerr;
218 }
219
220 if (IS_ERR(mirror->mirror_ds))
221 goto outerr;
222
223 if (mirror->mirror_ds->ds == NULL) {
224 struct nfs4_deviceid_node *devid;
225 devid = &mirror->mirror_ds->id_node;
226 ff_layout_mark_devid_invalid(lseg, devid);
227 return false;
228 }
229 return true;
230outerr:
231 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
232 return false;
233}
234
235static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
236 u64 offset, u64 length)
237{
238 u64 end;
239
240 end = max_t(u64, pnfs_end_offset(err->offset, err->length),
241 pnfs_end_offset(offset, length));
242 err->offset = min_t(u64, err->offset, offset);
243 err->length = end - err->offset;
244}
245
246static int
247ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
248 const struct nfs4_ff_layout_ds_err *e2)
249{
250 int ret;
251
252 if (e1->opnum != e2->opnum)
253 return e1->opnum < e2->opnum ? -1 : 1;
254 if (e1->status != e2->status)
255 return e1->status < e2->status ? -1 : 1;
256 ret = memcmp(e1->stateid.data, e2->stateid.data,
257 sizeof(e1->stateid.data));
258 if (ret != 0)
259 return ret;
260 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
261 if (ret != 0)
262 return ret;
263 if (pnfs_end_offset(e1->offset, e1->length) < e2->offset)
264 return -1;
265 if (e1->offset > pnfs_end_offset(e2->offset, e2->length))
266 return 1;
267 /* If ranges overlap or are contiguous, they are the same */
268 return 0;
269}
270
271static void
272ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
273 struct nfs4_ff_layout_ds_err *dserr)
274{
275 struct nfs4_ff_layout_ds_err *err, *tmp;
276 struct list_head *head = &flo->error_list;
277 int match;
278
279 /* Do insertion sort w/ merges */
280 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
281 match = ff_ds_error_match(err, dserr);
282 if (match < 0)
283 continue;
284 if (match > 0) {
285 /* Add entry "dserr" _before_ entry "err" */
286 head = &err->list;
287 break;
288 }
289 /* Entries match, so merge "err" into "dserr" */
290 extend_ds_error(dserr, err->offset, err->length);
291 list_replace(&err->list, &dserr->list);
292 kfree(err);
293 return;
294 }
295
296 list_add_tail(&dserr->list, head);
297}
298
299int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
300 struct nfs4_ff_layout_mirror *mirror, u64 offset,
301 u64 length, int status, enum nfs_opnum4 opnum,
302 gfp_t gfp_flags)
303{
304 struct nfs4_ff_layout_ds_err *dserr;
305
306 if (status == 0)
307 return 0;
308
309 if (mirror->mirror_ds == NULL)
310 return -EINVAL;
311
312 dserr = kmalloc(sizeof(*dserr), gfp_flags);
313 if (!dserr)
314 return -ENOMEM;
315
316 INIT_LIST_HEAD(&dserr->list);
317 dserr->offset = offset;
318 dserr->length = length;
319 dserr->status = status;
320 dserr->opnum = opnum;
321 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
322 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
323 NFS4_DEVICEID4_SIZE);
324
325 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
326 ff_layout_add_ds_error_locked(flo, dserr);
327 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
328
329 return 0;
330}
331
332static struct rpc_cred *
333ff_layout_get_mirror_cred(struct nfs4_ff_layout_mirror *mirror, u32 iomode)
334{
335 struct rpc_cred *cred, __rcu **pcred;
336
337 if (iomode == IOMODE_READ)
338 pcred = &mirror->ro_cred;
339 else
340 pcred = &mirror->rw_cred;
341
342 rcu_read_lock();
343 do {
344 cred = rcu_dereference(*pcred);
345 if (!cred)
346 break;
347
348 cred = get_rpccred_rcu(cred);
349 } while(!cred);
350 rcu_read_unlock();
351 return cred;
352}
353
354struct nfs_fh *
355nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
356{
357 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
358 struct nfs_fh *fh = NULL;
359
360 if (!ff_layout_mirror_valid(lseg, mirror, false)) {
361 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
362 __func__, mirror_idx);
363 goto out;
364 }
365
366 /* FIXME: For now assume there is only 1 version available for the DS */
367 fh = &mirror->fh_versions[0];
368out:
369 return fh;
370}
371
372/**
373 * nfs4_ff_layout_prepare_ds - prepare a DS connection for an RPC call
374 * @lseg: the layout segment we're operating on
375 * @ds_idx: index of the DS to use
376 * @fail_return: return layout on connect failure?
377 *
378 * Try to prepare a DS connection to accept an RPC call. This involves
379 * selecting a mirror to use and connecting the client to it if it's not
380 * already connected.
381 *
382 * Since we only need a single functioning mirror to satisfy a read, we don't
383 * want to return the layout if there is one. For writes though, any down
384 * mirror should result in a LAYOUTRETURN. @fail_return is how we distinguish
385 * between the two cases.
386 *
387 * Returns a pointer to a connected DS object on success or NULL on failure.
388 */
389struct nfs4_pnfs_ds *
390nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
391 bool fail_return)
392{
393 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
394 struct nfs4_pnfs_ds *ds = NULL;
395 struct nfs4_deviceid_node *devid;
396 struct inode *ino = lseg->pls_layout->plh_inode;
397 struct nfs_server *s = NFS_SERVER(ino);
398 unsigned int max_payload;
399 int status;
400
401 if (!ff_layout_mirror_valid(lseg, mirror, true)) {
402 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
403 __func__, ds_idx);
404 goto out;
405 }
406
407 devid = &mirror->mirror_ds->id_node;
408 if (ff_layout_test_devid_unavailable(devid))
409 goto out_fail;
410
411 ds = mirror->mirror_ds->ds;
412 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
413 smp_rmb();
414 if (ds->ds_clp)
415 goto out;
416
417 /* FIXME: For now we assume the server sent only one version of NFS
418 * to use for the DS.
419 */
420 status = nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
421 dataserver_retrans,
422 mirror->mirror_ds->ds_versions[0].version,
423 mirror->mirror_ds->ds_versions[0].minor_version);
424
425 /* connect success, check rsize/wsize limit */
426 if (!status) {
427 max_payload =
428 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
429 NULL);
430 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
431 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
432 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
433 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
434 goto out;
435 }
436out_fail:
437 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
438 mirror, lseg->pls_range.offset,
439 lseg->pls_range.length, NFS4ERR_NXIO,
440 OP_ILLEGAL, GFP_NOIO);
441 if (fail_return || !ff_layout_has_available_ds(lseg))
442 pnfs_error_mark_layout_for_return(ino, lseg);
443 ds = NULL;
444out:
445 return ds;
446}
447
448struct rpc_cred *
449ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
450 struct rpc_cred *mdscred)
451{
452 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
453 struct rpc_cred *cred;
454
455 if (mirror) {
456 cred = ff_layout_get_mirror_cred(mirror, lseg->pls_range.iomode);
457 if (!cred)
458 cred = get_rpccred(mdscred);
459 } else {
460 cred = get_rpccred(mdscred);
461 }
462 return cred;
463}
464
465/**
466* Find or create a DS rpc client with th MDS server rpc client auth flavor
467* in the nfs_client cl_ds_clients list.
468*/
469struct rpc_clnt *
470nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
471 struct nfs_client *ds_clp, struct inode *inode)
472{
473 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
474
475 switch (mirror->mirror_ds->ds_versions[0].version) {
476 case 3:
477 /* For NFSv3 DS, flavor is set when creating DS connections */
478 return ds_clp->cl_rpcclient;
479 case 4:
480 return nfs4_find_or_create_ds_client(ds_clp, inode);
481 default:
482 BUG();
483 }
484}
485
486void ff_layout_free_ds_ioerr(struct list_head *head)
487{
488 struct nfs4_ff_layout_ds_err *err;
489
490 while (!list_empty(head)) {
491 err = list_first_entry(head,
492 struct nfs4_ff_layout_ds_err,
493 list);
494 list_del(&err->list);
495 kfree(err);
496 }
497}
498
499/* called with inode i_lock held */
500int ff_layout_encode_ds_ioerr(struct xdr_stream *xdr, const struct list_head *head)
501{
502 struct nfs4_ff_layout_ds_err *err;
503 __be32 *p;
504
505 list_for_each_entry(err, head, list) {
506 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
507 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
508 * + status(4) + opnum(4)
509 */
510 p = xdr_reserve_space(xdr,
511 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
512 if (unlikely(!p))
513 return -ENOBUFS;
514 p = xdr_encode_hyper(p, err->offset);
515 p = xdr_encode_hyper(p, err->length);
516 p = xdr_encode_opaque_fixed(p, &err->stateid,
517 NFS4_STATEID_SIZE);
518 /* Encode 1 error */
519 *p++ = cpu_to_be32(1);
520 p = xdr_encode_opaque_fixed(p, &err->deviceid,
521 NFS4_DEVICEID4_SIZE);
522 *p++ = cpu_to_be32(err->status);
523 *p++ = cpu_to_be32(err->opnum);
524 dprintk("%s: offset %llu length %llu status %d op %d\n",
525 __func__, err->offset, err->length, err->status,
526 err->opnum);
527 }
528
529 return 0;
530}
531
532static
533unsigned int do_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
534 const struct pnfs_layout_range *range,
535 struct list_head *head,
536 unsigned int maxnum)
537{
538 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
539 struct inode *inode = lo->plh_inode;
540 struct nfs4_ff_layout_ds_err *err, *n;
541 unsigned int ret = 0;
542
543 spin_lock(&inode->i_lock);
544 list_for_each_entry_safe(err, n, &flo->error_list, list) {
545 if (!pnfs_is_range_intersecting(err->offset,
546 pnfs_end_offset(err->offset, err->length),
547 range->offset,
548 pnfs_end_offset(range->offset, range->length)))
549 continue;
550 if (!maxnum)
551 break;
552 list_move(&err->list, head);
553 maxnum--;
554 ret++;
555 }
556 spin_unlock(&inode->i_lock);
557 return ret;
558}
559
560unsigned int ff_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
561 const struct pnfs_layout_range *range,
562 struct list_head *head,
563 unsigned int maxnum)
564{
565 unsigned int ret;
566
567 ret = do_layout_fetch_ds_ioerr(lo, range, head, maxnum);
568 /* If we're over the max, discard all remaining entries */
569 if (ret == maxnum) {
570 LIST_HEAD(discard);
571 do_layout_fetch_ds_ioerr(lo, range, &discard, -1);
572 ff_layout_free_ds_ioerr(&discard);
573 }
574 return ret;
575}
576
577static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
578{
579 struct nfs4_ff_layout_mirror *mirror;
580 struct nfs4_deviceid_node *devid;
581 u32 idx;
582
583 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
584 mirror = FF_LAYOUT_COMP(lseg, idx);
585 if (mirror) {
586 if (!mirror->mirror_ds)
587 return true;
588 if (IS_ERR(mirror->mirror_ds))
589 continue;
590 devid = &mirror->mirror_ds->id_node;
591 if (!ff_layout_test_devid_unavailable(devid))
592 return true;
593 }
594 }
595
596 return false;
597}
598
599static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
600{
601 struct nfs4_ff_layout_mirror *mirror;
602 struct nfs4_deviceid_node *devid;
603 u32 idx;
604
605 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
606 mirror = FF_LAYOUT_COMP(lseg, idx);
607 if (!mirror || IS_ERR(mirror->mirror_ds))
608 return false;
609 if (!mirror->mirror_ds)
610 continue;
611 devid = &mirror->mirror_ds->id_node;
612 if (ff_layout_test_devid_unavailable(devid))
613 return false;
614 }
615
616 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
617}
618
619static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
620{
621 if (lseg->pls_range.iomode == IOMODE_READ)
622 return ff_read_layout_has_available_ds(lseg);
623 /* Note: RW layout needs all mirrors available */
624 return ff_rw_layout_has_available_ds(lseg);
625}
626
627bool ff_layout_avoid_mds_available_ds(struct pnfs_layout_segment *lseg)
628{
629 return ff_layout_no_fallback_to_mds(lseg) ||
630 ff_layout_has_available_ds(lseg);
631}
632
633bool ff_layout_avoid_read_on_rw(struct pnfs_layout_segment *lseg)
634{
635 return lseg->pls_range.iomode == IOMODE_RW &&
636 ff_layout_no_read_on_rw(lseg);
637}
638
639module_param(dataserver_retrans, uint, 0644);
640MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
641 "retries a request before it attempts further "
642 " recovery action.");
643module_param(dataserver_timeo, uint, 0644);
644MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
645 "NFSv4.1 client waits for a response from a "
646 " data server before it retries an NFS request.");
1/*
2 * Device operations for the pnfs nfs4 file layout driver.
3 *
4 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
5 *
6 * Tao Peng <bergwolf@primarydata.com>
7 */
8
9#include <linux/nfs_fs.h>
10#include <linux/vmalloc.h>
11#include <linux/module.h>
12#include <linux/sunrpc/addr.h>
13
14#include "../internal.h"
15#include "../nfs4session.h"
16#include "flexfilelayout.h"
17
18#define NFSDBG_FACILITY NFSDBG_PNFS_LD
19
20static unsigned int dataserver_timeo = NFS4_DEF_DS_TIMEO;
21static unsigned int dataserver_retrans = NFS4_DEF_DS_RETRANS;
22
23void nfs4_ff_layout_put_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
24{
25 if (mirror_ds)
26 nfs4_put_deviceid_node(&mirror_ds->id_node);
27}
28
29void nfs4_ff_layout_free_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
30{
31 nfs4_print_deviceid(&mirror_ds->id_node.deviceid);
32 nfs4_pnfs_ds_put(mirror_ds->ds);
33 kfree_rcu(mirror_ds, id_node.rcu);
34}
35
36/* Decode opaque device data and construct new_ds using it */
37struct nfs4_ff_layout_ds *
38nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
39 gfp_t gfp_flags)
40{
41 struct xdr_stream stream;
42 struct xdr_buf buf;
43 struct page *scratch;
44 struct list_head dsaddrs;
45 struct nfs4_pnfs_ds_addr *da;
46 struct nfs4_ff_layout_ds *new_ds = NULL;
47 struct nfs4_ff_ds_version *ds_versions = NULL;
48 u32 mp_count;
49 u32 version_count;
50 __be32 *p;
51 int i, ret = -ENOMEM;
52
53 /* set up xdr stream */
54 scratch = alloc_page(gfp_flags);
55 if (!scratch)
56 goto out_err;
57
58 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
59 if (!new_ds)
60 goto out_scratch;
61
62 nfs4_init_deviceid_node(&new_ds->id_node,
63 server,
64 &pdev->dev_id);
65 INIT_LIST_HEAD(&dsaddrs);
66
67 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
68 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
69
70 /* multipath count */
71 p = xdr_inline_decode(&stream, 4);
72 if (unlikely(!p))
73 goto out_err_drain_dsaddrs;
74 mp_count = be32_to_cpup(p);
75 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
76
77 for (i = 0; i < mp_count; i++) {
78 /* multipath ds */
79 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
80 &stream, gfp_flags);
81 if (da)
82 list_add_tail(&da->da_node, &dsaddrs);
83 }
84 if (list_empty(&dsaddrs)) {
85 dprintk("%s: no suitable DS addresses found\n",
86 __func__);
87 ret = -ENOMEDIUM;
88 goto out_err_drain_dsaddrs;
89 }
90
91 /* version count */
92 p = xdr_inline_decode(&stream, 4);
93 if (unlikely(!p))
94 goto out_err_drain_dsaddrs;
95 version_count = be32_to_cpup(p);
96 dprintk("%s: version count %d\n", __func__, version_count);
97
98 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
99 gfp_flags);
100 if (!ds_versions)
101 goto out_scratch;
102
103 for (i = 0; i < version_count; i++) {
104 /* 20 = version(4) + minor_version(4) + rsize(4) + wsize(4) +
105 * tightly_coupled(4) */
106 p = xdr_inline_decode(&stream, 20);
107 if (unlikely(!p))
108 goto out_err_drain_dsaddrs;
109 ds_versions[i].version = be32_to_cpup(p++);
110 ds_versions[i].minor_version = be32_to_cpup(p++);
111 ds_versions[i].rsize = nfs_block_size(be32_to_cpup(p++), NULL);
112 ds_versions[i].wsize = nfs_block_size(be32_to_cpup(p++), NULL);
113 ds_versions[i].tightly_coupled = be32_to_cpup(p);
114
115 if (ds_versions[i].rsize > NFS_MAX_FILE_IO_SIZE)
116 ds_versions[i].rsize = NFS_MAX_FILE_IO_SIZE;
117 if (ds_versions[i].wsize > NFS_MAX_FILE_IO_SIZE)
118 ds_versions[i].wsize = NFS_MAX_FILE_IO_SIZE;
119
120 if (ds_versions[i].version != 3 || ds_versions[i].minor_version != 0) {
121 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
122 i, ds_versions[i].version,
123 ds_versions[i].minor_version);
124 ret = -EPROTONOSUPPORT;
125 goto out_err_drain_dsaddrs;
126 }
127
128 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
129 __func__, i, ds_versions[i].version,
130 ds_versions[i].minor_version,
131 ds_versions[i].rsize,
132 ds_versions[i].wsize,
133 ds_versions[i].tightly_coupled);
134 }
135
136 new_ds->ds_versions = ds_versions;
137 new_ds->ds_versions_cnt = version_count;
138
139 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
140 if (!new_ds->ds)
141 goto out_err_drain_dsaddrs;
142
143 /* If DS was already in cache, free ds addrs */
144 while (!list_empty(&dsaddrs)) {
145 da = list_first_entry(&dsaddrs,
146 struct nfs4_pnfs_ds_addr,
147 da_node);
148 list_del_init(&da->da_node);
149 kfree(da->da_remotestr);
150 kfree(da);
151 }
152
153 __free_page(scratch);
154 return new_ds;
155
156out_err_drain_dsaddrs:
157 while (!list_empty(&dsaddrs)) {
158 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
159 da_node);
160 list_del_init(&da->da_node);
161 kfree(da->da_remotestr);
162 kfree(da);
163 }
164
165 kfree(ds_versions);
166out_scratch:
167 __free_page(scratch);
168out_err:
169 kfree(new_ds);
170
171 dprintk("%s ERROR: returning %d\n", __func__, ret);
172 return NULL;
173}
174
175static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
176 struct nfs4_deviceid_node *devid)
177{
178 nfs4_mark_deviceid_unavailable(devid);
179 if (!ff_layout_has_available_ds(lseg))
180 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
181 lseg);
182}
183
184static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
185 struct nfs4_ff_layout_mirror *mirror)
186{
187 if (mirror == NULL || mirror->mirror_ds == NULL) {
188 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
189 lseg);
190 return false;
191 }
192 if (mirror->mirror_ds->ds == NULL) {
193 struct nfs4_deviceid_node *devid;
194 devid = &mirror->mirror_ds->id_node;
195 ff_layout_mark_devid_invalid(lseg, devid);
196 return false;
197 }
198 return true;
199}
200
201static u64
202end_offset(u64 start, u64 len)
203{
204 u64 end;
205
206 end = start + len;
207 return end >= start ? end : NFS4_MAX_UINT64;
208}
209
210static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
211 u64 offset, u64 length)
212{
213 u64 end;
214
215 end = max_t(u64, end_offset(err->offset, err->length),
216 end_offset(offset, length));
217 err->offset = min_t(u64, err->offset, offset);
218 err->length = end - err->offset;
219}
220
221static int
222ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
223 const struct nfs4_ff_layout_ds_err *e2)
224{
225 int ret;
226
227 if (e1->opnum != e2->opnum)
228 return e1->opnum < e2->opnum ? -1 : 1;
229 if (e1->status != e2->status)
230 return e1->status < e2->status ? -1 : 1;
231 ret = memcmp(&e1->stateid, &e2->stateid, sizeof(e1->stateid));
232 if (ret != 0)
233 return ret;
234 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
235 if (ret != 0)
236 return ret;
237 if (end_offset(e1->offset, e1->length) < e2->offset)
238 return -1;
239 if (e1->offset > end_offset(e2->offset, e2->length))
240 return 1;
241 /* If ranges overlap or are contiguous, they are the same */
242 return 0;
243}
244
245static void
246ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
247 struct nfs4_ff_layout_ds_err *dserr)
248{
249 struct nfs4_ff_layout_ds_err *err, *tmp;
250 struct list_head *head = &flo->error_list;
251 int match;
252
253 /* Do insertion sort w/ merges */
254 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
255 match = ff_ds_error_match(err, dserr);
256 if (match < 0)
257 continue;
258 if (match > 0) {
259 /* Add entry "dserr" _before_ entry "err" */
260 head = &err->list;
261 break;
262 }
263 /* Entries match, so merge "err" into "dserr" */
264 extend_ds_error(dserr, err->offset, err->length);
265 list_del(&err->list);
266 kfree(err);
267 }
268
269 list_add_tail(&dserr->list, head);
270}
271
272int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
273 struct nfs4_ff_layout_mirror *mirror, u64 offset,
274 u64 length, int status, enum nfs_opnum4 opnum,
275 gfp_t gfp_flags)
276{
277 struct nfs4_ff_layout_ds_err *dserr;
278
279 if (status == 0)
280 return 0;
281
282 if (mirror->mirror_ds == NULL)
283 return -EINVAL;
284
285 dserr = kmalloc(sizeof(*dserr), gfp_flags);
286 if (!dserr)
287 return -ENOMEM;
288
289 INIT_LIST_HEAD(&dserr->list);
290 dserr->offset = offset;
291 dserr->length = length;
292 dserr->status = status;
293 dserr->opnum = opnum;
294 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
295 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
296 NFS4_DEVICEID4_SIZE);
297
298 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
299 ff_layout_add_ds_error_locked(flo, dserr);
300 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
301
302 return 0;
303}
304
305/* currently we only support AUTH_NONE and AUTH_SYS */
306static rpc_authflavor_t
307nfs4_ff_layout_choose_authflavor(struct nfs4_ff_layout_mirror *mirror)
308{
309 if (mirror->uid == (u32)-1)
310 return RPC_AUTH_NULL;
311 return RPC_AUTH_UNIX;
312}
313
314/* fetch cred for NFSv3 DS */
315static int ff_layout_update_mirror_cred(struct nfs4_ff_layout_mirror *mirror,
316 struct nfs4_pnfs_ds *ds)
317{
318 if (ds->ds_clp && !mirror->cred &&
319 mirror->mirror_ds->ds_versions[0].version == 3) {
320 struct rpc_auth *auth = ds->ds_clp->cl_rpcclient->cl_auth;
321 struct rpc_cred *cred;
322 struct auth_cred acred = {
323 .uid = make_kuid(&init_user_ns, mirror->uid),
324 .gid = make_kgid(&init_user_ns, mirror->gid),
325 };
326
327 /* AUTH_NULL ignores acred */
328 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
329 if (IS_ERR(cred)) {
330 dprintk("%s: lookup_cred failed with %ld\n",
331 __func__, PTR_ERR(cred));
332 return PTR_ERR(cred);
333 } else {
334 if (cmpxchg(&mirror->cred, NULL, cred))
335 put_rpccred(cred);
336 }
337 }
338 return 0;
339}
340
341struct nfs_fh *
342nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
343{
344 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
345 struct nfs_fh *fh = NULL;
346
347 if (!ff_layout_mirror_valid(lseg, mirror)) {
348 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
349 __func__, mirror_idx);
350 goto out;
351 }
352
353 /* FIXME: For now assume there is only 1 version available for the DS */
354 fh = &mirror->fh_versions[0];
355out:
356 return fh;
357}
358
359/* Upon return, either ds is connected, or ds is NULL */
360struct nfs4_pnfs_ds *
361nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
362 bool fail_return)
363{
364 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
365 struct nfs4_pnfs_ds *ds = NULL;
366 struct nfs4_deviceid_node *devid;
367 struct inode *ino = lseg->pls_layout->plh_inode;
368 struct nfs_server *s = NFS_SERVER(ino);
369 unsigned int max_payload;
370 rpc_authflavor_t flavor;
371
372 if (!ff_layout_mirror_valid(lseg, mirror)) {
373 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
374 __func__, ds_idx);
375 goto out;
376 }
377
378 devid = &mirror->mirror_ds->id_node;
379 if (ff_layout_test_devid_unavailable(devid))
380 goto out;
381
382 ds = mirror->mirror_ds->ds;
383 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
384 smp_rmb();
385 if (ds->ds_clp)
386 goto out_update_creds;
387
388 flavor = nfs4_ff_layout_choose_authflavor(mirror);
389
390 /* FIXME: For now we assume the server sent only one version of NFS
391 * to use for the DS.
392 */
393 nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
394 dataserver_retrans,
395 mirror->mirror_ds->ds_versions[0].version,
396 mirror->mirror_ds->ds_versions[0].minor_version,
397 flavor);
398
399 /* connect success, check rsize/wsize limit */
400 if (ds->ds_clp) {
401 max_payload =
402 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
403 NULL);
404 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
405 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
406 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
407 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
408 } else {
409 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
410 mirror, lseg->pls_range.offset,
411 lseg->pls_range.length, NFS4ERR_NXIO,
412 OP_ILLEGAL, GFP_NOIO);
413 if (!fail_return) {
414 if (ff_layout_has_available_ds(lseg))
415 set_bit(NFS_LAYOUT_RETURN_REQUESTED,
416 &lseg->pls_layout->plh_flags);
417 else
418 pnfs_error_mark_layout_for_return(ino, lseg);
419 } else
420 pnfs_error_mark_layout_for_return(ino, lseg);
421 ds = NULL;
422 goto out;
423 }
424out_update_creds:
425 if (ff_layout_update_mirror_cred(mirror, ds))
426 ds = NULL;
427out:
428 return ds;
429}
430
431struct rpc_cred *
432ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
433 struct rpc_cred *mdscred)
434{
435 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
436 struct rpc_cred *cred = ERR_PTR(-EINVAL);
437
438 if (!nfs4_ff_layout_prepare_ds(lseg, ds_idx, true))
439 goto out;
440
441 if (mirror && mirror->cred)
442 cred = mirror->cred;
443 else
444 cred = mdscred;
445out:
446 return cred;
447}
448
449/**
450* Find or create a DS rpc client with th MDS server rpc client auth flavor
451* in the nfs_client cl_ds_clients list.
452*/
453struct rpc_clnt *
454nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
455 struct nfs_client *ds_clp, struct inode *inode)
456{
457 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
458
459 switch (mirror->mirror_ds->ds_versions[0].version) {
460 case 3:
461 /* For NFSv3 DS, flavor is set when creating DS connections */
462 return ds_clp->cl_rpcclient;
463 case 4:
464 return nfs4_find_or_create_ds_client(ds_clp, inode);
465 default:
466 BUG();
467 }
468}
469
470static bool is_range_intersecting(u64 offset1, u64 length1,
471 u64 offset2, u64 length2)
472{
473 u64 end1 = end_offset(offset1, length1);
474 u64 end2 = end_offset(offset2, length2);
475
476 return (end1 == NFS4_MAX_UINT64 || end1 > offset2) &&
477 (end2 == NFS4_MAX_UINT64 || end2 > offset1);
478}
479
480/* called with inode i_lock held */
481int ff_layout_encode_ds_ioerr(struct nfs4_flexfile_layout *flo,
482 struct xdr_stream *xdr, int *count,
483 const struct pnfs_layout_range *range)
484{
485 struct nfs4_ff_layout_ds_err *err, *n;
486 __be32 *p;
487
488 list_for_each_entry_safe(err, n, &flo->error_list, list) {
489 if (!is_range_intersecting(err->offset, err->length,
490 range->offset, range->length))
491 continue;
492 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
493 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
494 * + status(4) + opnum(4)
495 */
496 p = xdr_reserve_space(xdr,
497 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
498 if (unlikely(!p))
499 return -ENOBUFS;
500 p = xdr_encode_hyper(p, err->offset);
501 p = xdr_encode_hyper(p, err->length);
502 p = xdr_encode_opaque_fixed(p, &err->stateid,
503 NFS4_STATEID_SIZE);
504 /* Encode 1 error */
505 *p++ = cpu_to_be32(1);
506 p = xdr_encode_opaque_fixed(p, &err->deviceid,
507 NFS4_DEVICEID4_SIZE);
508 *p++ = cpu_to_be32(err->status);
509 *p++ = cpu_to_be32(err->opnum);
510 *count += 1;
511 list_del(&err->list);
512 dprintk("%s: offset %llu length %llu status %d op %d count %d\n",
513 __func__, err->offset, err->length, err->status,
514 err->opnum, *count);
515 kfree(err);
516 }
517
518 return 0;
519}
520
521static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
522{
523 struct nfs4_ff_layout_mirror *mirror;
524 struct nfs4_deviceid_node *devid;
525 u32 idx;
526
527 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
528 mirror = FF_LAYOUT_COMP(lseg, idx);
529 if (mirror && mirror->mirror_ds) {
530 devid = &mirror->mirror_ds->id_node;
531 if (!ff_layout_test_devid_unavailable(devid))
532 return true;
533 }
534 }
535
536 return false;
537}
538
539static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
540{
541 struct nfs4_ff_layout_mirror *mirror;
542 struct nfs4_deviceid_node *devid;
543 u32 idx;
544
545 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
546 mirror = FF_LAYOUT_COMP(lseg, idx);
547 if (!mirror || !mirror->mirror_ds)
548 return false;
549 devid = &mirror->mirror_ds->id_node;
550 if (ff_layout_test_devid_unavailable(devid))
551 return false;
552 }
553
554 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
555}
556
557bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
558{
559 if (lseg->pls_range.iomode == IOMODE_READ)
560 return ff_read_layout_has_available_ds(lseg);
561 /* Note: RW layout needs all mirrors available */
562 return ff_rw_layout_has_available_ds(lseg);
563}
564
565module_param(dataserver_retrans, uint, 0644);
566MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
567 "retries a request before it attempts further "
568 " recovery action.");
569module_param(dataserver_timeo, uint, 0644);
570MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
571 "NFSv4.1 client waits for a response from a "
572 " data server before it retries an NFS request.");